Black Friday and Cyber Week, Fri 20 Nov to Tue 1 Dec: the AppIQ AI Suite, all ten AI products, for $19,997 (bought separately $29,970). See the Suite

AI Software / Proof Pack

AIProof Pack

Don't take our word for it. Test AI Bank Match yourself.

A public test set with the right answers written down in advance: a bank statement, a ledger and an answer key. Run it in the free example, mark every result against the key, then switch off your Wi-Fi and check that nothing leaves your computer. Every name in the files is invented.

The test set

Three files, data only

Plain CSV files that open in Excel or any text editor. Use them on AppIQ or on any other reconciliation software, and score both against the same key.

Bank statement

September 2026, 32 lines: date, description, amount, running balance.

Download CSV (1.9 KB)

SHA-256 fingerprint
4122944fe6ff651a551d9a1b02059a2488e43b374c13e5075a6664fd0674a18e

Ledger (bank account in the books)

September 2026, 30 lines: posting date, document number, narrative, debit, credit, balance.

Download CSV (2.2 KB)

SHA-256 fingerprint
df5d2a80d6515f16e202daeda157470922d155c6dae62b253d017e3f2d66186f

Answer key

All 32 items with the right answer written down before the test: which lines belong together, and which must stay apart.

Download CSV (5.9 KB)

SHA-256 fingerprint
f0b8c9cd4814e2987f4e30fe87b10cbc2c5cf581e0ebcbcb9724b7439a02c560

What is in it

32 items, each one a known answer

Right answerItemsWhat they test
Should be matched23Same amount and invoice number; same amount and payee; two payments of the same amount in the same week where only the payee decides; one deposit of three receipts; one payroll line paid in three batches; payees worded differently on each side (telecom, water, courier, fuel, flights, taxi, staff lunch); an invoice posted twice but paid once.
Must NOT be matched automatically2The same payment with two digits swapped (1,650.00 against 1,560.00), and a refund posted the wrong way round.
Only on the bank statement5Two bank charges, interest, money received that nobody has posted, and a planted round payment on a Sunday to a payee the ledger has never seen.
Only in the ledger2An unpresented cheque and a deposit in transit.

In the answer key, line numbers count the rows under the header: bank line 1 is the first transaction in the bank file.

Run it

Two minutes in the free example

  1. Open the free example of AI Bank Match in Chrome or Edge on a computer. The test set loads by itself. (On the page, it is the Load the public test set link under step 1.)
  2. Check it is the same data. Step 2 shows each file's totals. They should read: Bank statement: 32 lines, money in 22,909.62, money out -85,026.10, and Ledger: 30 lines, money in 22,285.50, money out -79,552.35. These are the totals of the downloaded files.
  3. Leave the settings as they are (lines up to 7 days apart, combined lines on, on-device AI on) and press Match now. The first time, the AI model (about 34 MB) downloads from appiqfinance.com.
  4. Open each tab (Matched, To check, Only on bank, Only in ledger, Likely reasons) and mark every line against the answer key.
Expected result

What you should see

We ran the product's own matching engine and AI model on these files, then ran the test set again in the product page itself in the browser, and got the same result. An automated test in our test suite re-checks this result. Computed 12 November 2026.

17Matched
6To check
7Only on bank
5Only in ledger
6Likely reasons
0Wrong automatic matches
Every pair is right

All 17 matches agree with the answer key, and each of the 6 pairs in To check is a right pair waiting for your yes. That is all 23 items that should be matched: 17 matched automatically, 6 for you to confirm, none paired wrongly.

Lines that must stay apart, stay apart

The swapped digits and the wrong-way-round refund are not matched. They are listed under Likely reasons, with the duplicate invoice posting and the three bank charges and interest.

The reconciliation ties

With the closing balances read from the files (bank 22,883.52, ledger 27,733.15), both adjusted balances come to 21,053.52: difference 0.00.

With the AI switched off

Untick the AI in the settings, or go offline before the model has downloaded, and the rules alone give 16 matched, 5 to check, 9 only on bank and 7 only in ledger, again with no wrong match. On this set the AI adds 2 pairs: the 315.00 taxi fare (matched) and the 315.00 staff lunch (to check).

What the AI does not settle here: the telecom, water, courier, fuel and flights payments stay in To check. The names are invented, so the model has only ordinary words to go on, and AppIQ sends a pair it is not sure of to a person rather than guess.

How the AI products are tested

The wider test suite

Known answers, every match checked

Test generators build bank statements, ledgers and other files with a recorded right answer for every line. Each match the engine makes is compared with that answer; a single wrong automatic match fails the test.

0 wrong automatic matches on 40 random data sets

40 random bank data sets of 300 lines each: no wrong automatic match, and 100.00% of the ledger lines that are on the bank statement matched. Re-run on 12 November 2026 with the same figures.

The AI counted honestly

On 3,311 bank and ledger pairs worded differently, the rules alone matched 1,189 and the rules with the on-device AI 2,755, with no wrong High or Medium match. Re-run on 12 November 2026 with the same figures.

Test it with Wi-Fi off

Your browser blocks the page from sending anything

The AI Bank Match page is served with a security rule, a Content-Security-Policy header, that your browser enforces. connect-src 'none' means the page may not open a connection to any server, ours included. form-action 'none' means it may not submit a form. Scripts may come only from appiqfinance.com itself. The full rule:

default-src 'none'; script-src 'self' 'wasm-unsafe-eval'; style-src 'self'; font-src 'self'; img-src 'self' data: blob:; connect-src 'none'; form-action 'none'; frame-ancestors 'none'; base-uri 'none'; object-src 'none'
1. Read the rule

On the free example page press F12 (Mac: Cmd + Option + I) and open the Network tab. Reload the page. Click the first row, ai-match-app, then Headers. Under Response Headers, content-security-policy includes connect-src 'none'.

2. Watch the traffic

In the Network tab choose the Fetch/XHR filter, then load the test set and press Match now. The list stays empty. Under All you will see only files downloaded from appiqfinance.com: the page's own scripts and, the first time, the AI model in parts. Nothing goes out with your files.

3. Try to break it

Open the Console tab, type fetch("https://example.com") and press Enter. The browser refuses with a red message that names connect-src 'none'. If the page can't reach example.com, it can't reach anywhere else either.

4. Switch the Wi-Fi off

Run the match once with Wi-Fi on so the AI model is in your browser. Then switch off Wi-Fi (or unplug the network cable), but don't reload the page. Click Load the public test set again and press Match now: the same result appears, with no connection at all.

Questions

Why publish a test set?

Finance software should be checked like a control. With the right answers written down in advance, anyone can run the files and mark every result right or wrong, without having to trust us.

Is any of this real data?

No. Every company name, amount and reference in the files is invented. No customer data and no real people.

Can I load the downloaded files into the free example myself?

The free example runs only built-in files; your own files need an AI Bank Match Pro key. Its 'Load the public test set' link loads a copy that is byte for byte the same as the downloads: step 2 shows the same line counts and totals, and an automated test in our test suite fails if the copies ever differ. With a Pro key you can load the downloaded files themselves, and the 7-day money-back guarantee covers that test too.

Can I run the test set on other software?

Yes. The files are plain CSV that any reconciliation tool or spreadsheet can read, so you can score another product against the same answer key.

Will I get the same result on my own files?

Not necessarily. Real files differ in wording, volume and quality. What the test set shows is how the engine behaves: it pairs only what it can support, sends doubtful pairs to To check, keeps lines apart that must not be matched, and the reconciliation ties to the cent.

Why does the AI add only two pairs here?

Every name in the test set is invented, so the on-device language model has only the ordinary words in each description to go on, such as cabs, telecom or fuel. Pairs it is not sure of stay in To check for a person to confirm. The separate test on 3,311 differently worded pairs is on the test results page.